ESSAYS ON SOLAR PHOTOVOLTAIC ADOPTION AND ELECTRICITY CONSUMPTION PATTERNS: EVIDENCE FROM PARADISE A DISSERTATION SUBMITTED TO THE GRADUATE DIVISION OF THE UNIVERSITY OF HAWAIʻI AT MĀNOA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF PHD IN ECONOMICS AUGUST 2017 By Chasuta Anukoolthamchote Dissertation Committee: Denise Konan, Chairperson Lee Endress Timothy Halliday Nori Tarui Anthony Kuh Dora Nakafuji Keywords: solar PV adoption, rooftop PV, PV penetration, electricity consumption behavior, electricity demand, renewable energy, technology diffusion
124
Embed
Chasuta Anukoolthamchote - University of Hawaii€¦ · electricity load, specifically in light of increasing levels of solar saturation in Oahu. It is found that, regardless of the
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
ESSAYS ON SOLAR PHOTOVOLTAIC ADOPTION AND ELECTRICITY
CONSUMPTION PATTERNS: EVIDENCE FROM PARADISE
A DISSERTATION SUBMITTED TO THE GRADUATE DIVISION OF THE
UNIVERSITY OF HAWAIʻI AT MĀNOA IN PARTIAL FULFILLMENT OF
THE REQUIREMENTS FOR THE DEGREE OF
PHD
IN
ECONOMICS
AUGUST 2017
By
Chasuta Anukoolthamchote
Dissertation Committee:
Denise Konan, Chairperson
Lee Endress
Timothy Halliday
Nori Tarui
Anthony Kuh
Dora Nakafuji
Keywords: solar PV adoption, rooftop PV, PV penetration, electricity consumption
H.8 Percentage Change from Pre-Solar Usage – Separated by Percent Energy Offset. . . . . 99
1
CHAPTER 1
Increasing Solar PV Penetration
& Fluctuation in Net Electricity Consumption
1.1 Introduction
Achieving adequate supplies of clean energy for the future is a great societal challenge. Nowhere
is this need more urgent than in Hawaiʻi, where electricity prices are three times higher than the
U.S. mainland average.1 These high energy costs have a large impact on Hawaiʻi’s economy,
imposing a major burden on both local customers and businesses. As the demand for energy
continues to grow, Hawaiʻi has focused on transitioning to clean and affordable renewable
energy sources.
Solar photovoltaic (PV) adoptions have grown exponentially within Hawaiʻi. Figure B.1
illustrates annual and cumulative installations of solar PV by customer segment on the island of
Oahu.2 However, the widespread adoption of solar PV poses a number of unique challenges to
electric grids which must integrate large quantities of solar generation. Solar output driven by
solar irradiance is variable and intermittent, and cannot be adjusted by the utility system
operator. Rapid swings in solar electricity can lead to temporary mismatches between energy
supply and demand. Therefore, additional dispatchable system reserves and backup capacity may
be necessary in order to maintain system reliability.
This study evaluates the impact of intermittent solar power generation under variable climate
conditions on the electric grid of the island of Oahu from September 2010 to May 2014. The
standard deviation of net electricity load is used as a representative variable for volatility in
electricity generated by the utility.3 An empirical analysis is performed employing a unique
proprietary data set detailing electricity net load and solar penetration levels at the distribution
1 For example, in June 2014, Hawaiʻi’s electricity price was 38.7 cents per kilowatt-hour (kWh) while the U.S
Mainland average was 12.9 cents per kWh. 2 The megawatt installed capacity shown in figure B.1 is measured by the total "nameplate" capacity of solar PV of
fully executed applications. 3 Net electricity load is defined as the amount of energy met by utility generation.
2
transformer level.4 It is found that higher levels of PV penetration increase variability of net load.
This effect is most pronounced during daytime hours when the sun is out.
The data set employed in this study also contains detailed information on the customer mix of
each distribution transformer. Using this information, we examine the combined effect of
customer diversity and increased PV penetration on the shape of load profiles. It is found that,
regardless of the level of solar power generation, customer mix exerts a significant impact on the
pattern and level of net load from hour to hour. In residential-concentrated areas, net load
exhibits two distinct peaks – morning and night. This is reflective of the underlying consumption
behavior of residential customers who typically leave for school or work in the morning and
return in the evening. Conversely, commercial- and industrial-concentrated areas exhibit a single
midday peak corresponding to regular business hours.
The different distribution transformers considered in this study are observed to have a diversity
of load patterns, depending upon their mix of customer classes. While the shape of their load
profiles is primarily influenced by the customer mix and time of day, their load variability at
each time period is remarkably the result of increased PV penetration level. Our result indicates
that customer mix is the key driver influencing net load behavior in each area. Taking into
account the increased PV adoptions, the dynamic between residential and commercial electricity
use patterns immensely reduces issues resulting from high variability in solar power generation.
Over the past few years, Hawaiian Electric Company (HECO) has deployed several solar and
wind monitoring devices at various locations throughout its service territory. These tools provide
estimates of solar irradiance, wind speed, as well as temperature and relative humidity. In this
study, we employ these measured meteorological variables along with seasonal variations to
identify both the impact of climatic variations on electricity use, as well as their impact upon
intermittent solar power generation. We find that wind speed negatively affects net load
variability, whereas temperature exhibits a significant and positive impact. Given the semi-
homogenous climate in Oahu, we further investigate the effect of temperature while taking into
account humidity values in the air. Our results show that when humidity is high, increasing
4 A distribution transformer provides the final voltage transformation in the electric power distribution system,
changing voltage level between higher transmission voltages and lower distribution voltages. In the sample of this
paper, there are one to three distribution transformers in a substation depending upon the size of cities or towns
which the substation supplies power to.
3
temperatures cause net load variability to increase at a diminishing rate. This is largely due to
consumers’ electricity consumption behaviors. Most consumers consume more energy by turning
on cooling electrical appliances when the temperature rises. This effect persists when both
temperature and humidity are high. The effect of increases in both factors, however, increases
the volatility in net load at a decreasing rate.
The seasonal variations are captured through time-of-day and season dummy variables. Our
results indicate that the net load exhibits higher volatility during the day in winter months.
During the winter, Hawaiʻi’s temperature and solar radiation levels are generally lower than
those experienced in the summer. Changes in temperature between day and night will likely
induce consumers to switch on air conditioning during the day and off during the evening.
Moreover, the solar output will likely fluctuate more due to the high intermittency of solar
resources in the winter, resulting in larger volatility in net electricity load.
The remainder of the paper is organized as follows. Section 1.2 reviews related literature. In
Section 1.3, we introduce the details of our unique data set and describe how each variable is
calculated. Section 1.4 then presents the econometric models used in this analysis. Estimation
results are reported in Section 1.5. The concluding remarks and discussion are given in section
1.6.
1.2 Literature Review
Several studies address the value and impact of the intermittency of solar electricity generation
upon electric grids (Hansen, 2007; Fthenakis et al., 2009; Joskoaw, 2011; Stein et al., 2012;
Stewart et al., 2013). These studies claim that, in addition to certain characteristics of PV panels,
various other factors affect solar electricity output, namely, the time of day, change of seasons
and weather variations. As examples, Mills (2013) and Baker et al. (2013) examined the
economic value of variable renewable generation and concluded that the value of new
generations of PV can decrease as the level of penetration rises. Because there are additional
costs associated with backup generation and solar intermittency, such costs can be minimized if
utility companies know how to better operate the electric grid. If operations and investments are
optimally managed in consideration of PV penetration, then additional costs can be
comparatively small.
4
Consequently, forecasts of solar power have become necessary for the integration of fluctuating
renewable energy into the grid. Various emerging studies have considered the ability to
accurately forecast variability in renewable resources, such as wind and solar (Perez et al., 2010;
Lorenz et al., 2011; Marquez and Coimbra, 2011; Mathiesen and Kleissl, 2011). Because solar
power typically exhibits different generation characteristics as compared with power produced
by conventional sources, more precise solar forecasts will enable electric system operators to
better manage electricity generation in spite of fluctuating solar output.
Along with the volatility of solar output, fluctuations in electricity usage also depend upon
climatic conditions. The influence of weather variations has a demand side impact on the
electricity market. Weather can have diverse effects on different sectors of the economy. Moral-
Carcedo and Pérez-García (2015), for example, find that changes in temperature have a large
effect on the electricity demand in Spain’s service sector. On the other hand, weather variations
have not been found to influence activities of the industrial sector, remaining highly and
positively correlated with residential electricity demand (Amato et al., 2005; Zachariadis and
Pashourtidou, 2007; Asadoorian et al., 2008; Bessec and Fouquau, 2008; Vassileva et al., 2012;
Ahmed et al., 2012). In other words, the effect of climate on electricity demand depends largely
on the main use of electricity, as influenced by various weather characteristics (Lam et al., 2008).
Although temperature is widely known to be highly correlated with electricity consumption, it is
not the only climatic variable considered in the literature. Sailor and Muñoz (1997); Yan (1998);
Valor et al. (2001); Hor et al. (2005); Hekkenberg et al. (2009); Apadula et al. (2012); Tung et al.
(2013) include temperature data along with other weather variables such as relative humidity,
wind speed, cloud cover, and solar radiation to evaluate the impact of climate on electricity
consumption. They conclude that temperature is the most significant climatic factor.
1.3 Data Description & Summary Statistics
The primary dataset used in this study was provided by HECO. The dataset was made available
to the University of Hawaiʻi Research Organization (UHERO) under a confidentiality agreement.
Covering a period from September 2010 to May 2014, the data contains measurements of net
electricity load, solar PV capacity installed, a number of customers on each rate schedule, and a
variety of weather variables including average temperature, relative humidity, average wind
speed and solar irradiance for 115 distribution transformers on Oahu, Hawaiʻi.
5
For the purpose of this study, electricity net load is defined to be the amount of energy met by
utility generation, while electricity gross load corresponds to the total demand, or total electricity
used, by customers. Electricity gross load is met by a combination of electricity provided by the
utility and electricity produced by local distributed generations such as wind and rooftop PV
systems. The difference between net and gross load is, therefore, the amount of electricity
produced by distributed generations. Without the precise measure of behind-the-meter solar
output, gross load or the total electricity demand is rather difficult to accurately estimate. As a
result, we will limit the focus of our study to the impact of increased PV penetration on net
electricity load.
Table A.1 provides summary statistics of variables used in this analysis at a transformer level.
The net load data analyzed in this paper consists of load measurements taken every 15 minutes,
corresponding to a total of 96 daily recorded values, and covering the period from September
2010 to May 2014. Using this data, the representative daily load set for different months and
years in the study period can be derived. For every month, the net load values recorded at a given
time of day (e.g., 2:15 or 14:45 hours) are used to calculate the standard deviation (SD) for each
of the 96 daily time periods. For example, the SD of all load values recorded at the hour of 12:30
from June 1, 2011 through June 30, 2011 are used to calculate the overall SD at the hour of
12:30 for the month of June 2011. These SD values are considered to be representative load
variabilities for the given month. Table A.1 shows standard deviation of net load ranges from 3.1
kW to 1800.5 kW with the mean of 270.9 kW.5
The standard deviation of net load calculated for each 15-minute period captures the volatility of
“net” electricity consumption. The shape of a given areas’ load profile is most influenced by its
customer-mix and the time-of-day, while its load volatility is largely the result of weather
patterns and the level of PV penetration. We describe summary statistics and how we calculate
these variables in the following sections.
1.3.1 Time-of-Day & Customer Mix
The sample data covers 115 distribution transformers servicing 199,704 distinct customers.
Table A.1 shows summary statistics of the number of residential and commercial customers on
5 Outliers and errors are eliminated by checking daily load graphs. We deleted the whole day that contains errors or
outliers. Holidays are also excluded.
6
each transformer. Of the 199,704 total customers contained within the dataset, 21,381 are
classified as commercial while the remaining 178,323 are classified as residential. The customer
classification is based on their rate schedules.6 The number of customers on a single transformer
ranges between 45 and 5,459 depending on the location where the distribution transformer
supplies power to. Although the number of customers significantly differs across areas and varies
over time, the data on the number of customers in this study stays constant across time, due to
the limited availability of historical information. We use the data on the total number of
customers and customers on residential rate schedule to calculate the percentage of residential to
total customers within each area, which ranges from 0% to 99% and averages 79%.
Time-of-Day
Table A.2 extends summary statistics for the standard deviation of net load by time-of-day,
season and year. It is observed that the statistics of net load is higher during the day and
increasing year-over-year. As expected, the average of net load volatility is greater during the
day due to both the consumption behavior of customers and the nature of intermittent solar
power generated by rooftop PV. An increase in average net load variations over the years
suggests that increasing penetration of PV makes net load more volatile. Between each season,
however, net load volatility behaves very similarly.
Customer Mix
Different transformers in our data set are observed to exhibit a variety of shapes and load profile
patterns. Figures B.2, B.3, and B.4 illustrate representative daily load patterns during each year
for three different customer-mixes with relatively high percentage of PV penetration. These
figures depict both the annual average net load profile (black solid line) and the load volatility
(green band represents the maximum and minimum values) throughout the day, and the annual
kilowatt (kW) PV installed capacity (solid red line). It is observed that customer-mix exerts a
significant impact on the pattern and level of net load from hour-to-hour.
6 Note that HECO currently does not have a separate category for industrial customers. Particularly, customers are
classified into two main types: residential and commercial. Commercial customers are further subdivided into three
major classes: small, medium, and large commercial dependent upon the magnitude of their electricity use. In this
paper, we use the detail of customer information on each transformer to determine whether customers are industrial
or commercial.
7
Figure B.2 is an example of a residential concentrated area. This transformer consists of 99%
residential households and only a few small commercial customers. When examining the annual
average net load in 2011 (solid black line), two distinct peaks are noted. The first peak
corresponds to morning (5:00 am - 7:00 am) and the second peak to evening (6:00 pm - 8:00
pm). This pattern can be attributed to residents waking up in the morning and later returning
following a day at work or school. Between the peaks, the net load is observed to fall during the
midday when the majority of customers are outside of their homes.
Figures B.3 and B.4 show representative annual average daily net load profiles for commercial-
and industrial-concentrated areas between 2011 and 2014, respectively. Commercial and
industrial load profiles often display similarities depending on the type of commercial and
industrial customers on the transformers. In 2011, before large gains had been made in solar PV
adoption, both load profiles exhibited a midday peak corresponding to regular business hours,
which gradually decreased towards a nighttime low as the majority of businesses closed for the
day. Although the representative commercial transformer captured in figure B.3 is comprised
only of 19% commercial customers, they are majority medium and large commercial users.7
Conversely, the industrial-concentrated population in figure B.4 consists of 100% commercial
and industrial customers.
Although the load profiles for commercial- and industrial-concentrated areas exhibit similar
patterns during weekdays, they are found to diverge from one another during the weekend.
Figure B.5, which displays 7-day net load profiles from August 28, 2011 to September 3, 2011,
illustrates the difference in load patterns between the three primary customer-mixes transformers.
It is observed that the load profile of the sample commercial-concentrated transformer is largely
uniform throughout the week, with little deviation between weekdays and weekends.
Conversely, the industrial-concentrated transformer exhibits divergent load profiles between
weekdays and weekends. Weekday load profiles of industrial-concentrated transformers typically
mimic those of commercial-concentrated ones, exhibiting a midday peak followed by a nighttime
lull. However, average weekend loads are observed to remain flat at minimum usage levels. This
7 Note that medium and large commercial customers consume 28.9% and 41.6% of the total annual electricity
consumption on Oahu. Since these commercial customers draw a large amount of electricity on this transformer,
having only 19% of commercial customers is sufficient to convey a representative commercial load profile.
8
pattern can be attributed to the fact that industrial customers do not typically operate during
weekends, although certain machines and/or electrical appliances may remain on when
businesses are closed. This differs from weekend loads of commercial-concentrated areas where
commercial customers such as groceries and department stores remain open throughout the
weekend, leading to load profiles that exhibit little variation over the course of a typical week.
Several studies Apadula et al. (2012); Hekkenberg et al. (2009); Pardo et al. (2002); Moral-
Carcedo and Vicens-Otero (2005) have shown that daily electricity demand is strongly
influenced by calendar effects, repeated sequences of weekdays and weekends which constitute
underlying periodic 7-day trends. These weekly cycles are among the main drivers of short-term
variations in electricity use. The recurring trend of higher electricity consumption during
weekdays is common to industrial-concentrated areas. Electricity consumption patterns of
residential households, wherein midday consumption is higher on weekends when a majority of
households are home, are also displayed in figure B.5.
1.3.2 Weather Variations & Solar PV Penetration
Weather data from a variety of sources is leveraged in support of this study. Weather variables
are estimated by averaging 15-minute time intervals in the same manner as was used for net load
data. Measurements of air temperature, relative humidity, and average wind speed were recorded
by weather sensors installed in several different areas on the island of Oahu. Weather variables
were queried on the basis of their relative geographic location to transformers and in the form of
power per unit area. Average temperature and relative humidity data were gathered by sensors at
four distinct locations on Oahu. By matching each transformer to its closest sensor, an estimate
of weather variations in each area may be attained. Average wind speed data was sourced from
two wind farms located on the northern coast of Oahu and is assumed to be identical for all
transformers in the study.
A number of other studies employed temperature-derived variables such as cooling (CDD) and
heating degree-days (HDD) when examining the impact of weather on electricity demand (Valor
et al., 2001; Hor et al., 2005; Amato et al., 2005; Moral-Carcedo and Vicens-Otero, 2005;
Zachariadis and Pashourtidou, 2007; Ahmed et al., 2012; Blázquez et al., 2013). Other studies
elected to exploit a wide combination of weather variables including temperature, humidity,
wind speed, cloud cover, rainfall, and solar radiation. (Engle et al., 1986; Filippini, 1995; Henley
9
and Peirson, 1997; Sailor and Muñoz, 1997; Yan, 1998; Henley and Peirson, 1998; Considine,
2000; Valor et al., 2001; Pardo et al., 2002).
In this study, a number of climatic variables are utilized to better understand the effect of
weather on the volatility of electricity load. Climatic variables were used in lieu of CDD/HDD as
an indicator of weather variability in this paper for a number of reasons. First, HDD data are
always zero in Oahu, Hawaiʻi. Although monthly CDD data are non-zero, the use of more
granular data is better than the use of a daily measurement of CDD. Leading to a second
justification, the frequency of our weather data is at the 15-minute time interval. The higher
frequency of our variables in the data set enhances the analysis at each 15 minute.
Temperature
Hawaiʻi exhibits considerably less temperature variation compared to other states. From table
A.1, the mean average temperature is 24.36 degrees Celsius with a standard deviation of 3.11
degrees. The highest and lowest average temperatures were observed in September and February,
respectively.
Relative Humidity
Relative humidity is a measure of moisture in the air which plays an important role in how
people perceive temperature.8 Sweat evaporates easily when the relative humidity is low, cooling
the body. Conversely, when relative humidity is high, sweat evaporates less readily leading to
higher perceived temperatures. At the extreme, when relative humidity reaches 100%, sweat no
longer evaporates into the air.9 The effect of both air temperature and humidity plays a major
role in a person’s likelihood to utilize air conditioning. Given the relatively small variations in
temperature experienced in Hawaiʻi, the additional consideration of humidity provides a better
understanding of how customers perceive and respond to different air temperature values.
The summary statistics of relative humidity are shown in table A.1 with a mean of 71.19% and a
standard deviation of 9.59%. The highest relative humidity during the study period was recorded
8 Relative humidity is a percentage of the maximum amount of water vapor that the air could hold at a given
temperature. 100% relative humidity implies that the air is totally saturated with water vapor and cannot hold
anymore, creating the possibility of rain. However, the relative humidity near the ground is generally much less than
100%. 9 For example, if the air temperature is 25 degree Celsius and the relative humidity is 0%, the air temperature feels
like 22 degree Celsius to our bodies. But when the relative humidity is 100%, we feel like it is 28 degree Celsius.
10
in August where the lowest was in January. Figure B.6 depicts a negative relationship between
average temperature and humidity, whereby humidity is lower when the temperature is high.
This negative relationship is most pronounced when solar is greater than zero (orange dots),
particularly during periods of high sunlight. This is intuitive given that sunlight reduces water
vapor in the air, lowering humidity and increasing temperature.
Wind Speed
Recorded average wind speed ranges from 3 to 13 miles per hour (mph), with a mean of 7.79
mph and a standard deviation of 2.59 mph. The highest and lowest average wind speeds typically
occurred during the months of July and January, respectively. Wind patterns on Hawaiʻi are
strongly influenced by the trade winds. Average wind speed is estimated for 15-minute time
intervals from data recorded at two different wind farms located on the northern coast of Oahu.
Given the limited number of wind speed sensors, there is considerably less variation in wind
speed data among transformers in the sample.
Solar Irradiance
Solar irradiance is a measure of power per unit area produced by the sun in the form of
electromagnetic radiation. Solar irradiance data in this study was calculated from PV power
output provided by anonymous customers with rooftop PV systems. Each customer is first
mapped to a transformer on the basis of their geographic location. The average solar power
produced by residential rooftop PV systems within each area is calculated in order to construct
an overall solar profile. This study utilizes a normalized measure of solar irradiance with values
falling between 0 and 1. Solar irradiance was found to range between 0 and 0.89, with an average
of 0.19 and a standard deviation of 0.26 (see table A.1). Figure B.7 illustrates a similar pattern of
average solar irradiance in winter and summer months at each time period throughout the day.
PV Penetration Level
The percentage of PV penetration reflects the percentage of daytime minimum electricity
consumption on a transformer that is covered and/or generated by local rooftop PV systems. In
this study, we employ percentage of PV penetration as a proxy for the amount of solar saturation
on each transformer. It can be calculated as follows:
11
% of PV Penetration = PV Installed Capacity (kW) x 100 (1.1)
Daytime Minimum Load (kVA)
where daytime minimum load (DML) represents the minimum load for a given location during
daytime (9:00 am to 5:00 pm). We calculate the value of DML of each transformer by first
matching each individual distribution circuit to their corresponding transformer.10 The DML of
each circuit under a given transformer is then summed to determine the DML at the transformer
level. Table A.1 shows that DML ranges between 108 and 4,412.2 kVA with a mean of 2,230.39
and a standard deviation of 896.05. The dataset detailing the amount of solar PV capacity
installed on each transformer was provided by HECO. It consists only of those PV customers
with executed agreements.11 The dataset provides details on solar system size, or PV nameplate
capacity, along with the date of installation.
The electricity load of each distribution transformer not only behaves differently throughout the
course of the day but also changes dramatically as solar PV penetration rises. In the sample
residential transformer shown in figure B.2, a clear decrease in daytime net load is observed
from 2011 to 2014. This is due in large part to a significant increase in the level of PV
penetration. In 2011, when the number of PV installations was relatively low, the annual net load
profile showed an increasing load during the midday with a nighttime peak. By 2012, following
an increase in PV installations, the midday load had dropped while simultaneously
demonstrating higher volatility during daytime hours when the sun was out and solar power was
readily available. Beginning in 2014, a back-feed problem on the residential-concentrated
transformer is observed, with the minimum net load declining below zero. During daytime hours,
the combination of lower residential power demand and higher electricity generated from rooftop
PV causes the net load profile to drop substantially. The nighttime peak does not, however,
exhibit the same degree of change over the study period. This implies that the nighttime behavior
of residential consumers remained largely unchanged despite the rise in PV installation.
10 Each distribution transformer in this study consists of one to four distribution circuits depending on the number of
customers, their relative location, and grid infrastructure. 11 These agreements include PV customers under Net Energy Metering, Feed-in Tariff, and Standard
Interconnection Agreement.
12
In addition, it can be clearly observed that in figure B.2 the midday sag in residential energy
demand becomes more pronounced from 2011 to 2014 as rooftop PV energy generation exceeds
energy demand. During this time period, the midday variation is also observed to increase. The
steep curve following the midday low rises as solar energy diminishes and late afternoon
consumption rises. There is a marked disruption attributable to PV generation, which causes net
load to behave differently than before. Due to this, the utility is confronted with a situation
wherein it must turn down its generators when solar reaches its peak, and then later ramp them
up more quickly than usual when solar power availability declines in the evening.
Figure B.4, a representative industrial-concentrated transformer, displays a dramatic decrease in
daytime net load. As discussed earlier, industrial customers typically use less electricity on
weekends when businesses are closed. The lower band, representing the minimum net load is
observed to drop below zero starting in 2012 due to a sharp increase in the number of solar PV
installations. The variation in net load increases over the course of the study period due to excess
electricity generated by rooftop PV, especially during weekends when demand is at its minimum.
This back-feeding during weekends lowers the average net load on the transformer over time.
1.4 Methodology
In line with the availability and description of our data in the previous section, we employ a fixed
effects model, which controls for time-of-day, month and year effects, and a set of covariates.
Our goal is to capture the effect of relevant variables on net load volatility. We hypothesize that
the volatility of net electricity load at time t for each distribution transformer i depends on the
level of PV penetration, the mix of customer classes, weather variations, time-of-day and
where TempHumit is the interaction term between temperature and humidity. As mentioned in the
previous section, with the relatively small variations in temperature in Hawaiʻi, the level of
15
humidity in the air can make it feel hotter or cooler. This interaction effect is used to capture the
impact of both temperature and humidity on fluctuations in net load.
1.5 Empirical Results
Results of several specifications are reported in table A.3. In addition to the models specified in
the previous section, we include models of two or more interaction terms to capture the effects of
explanatory variables. Column (1) in table A.3 reports regression results of our baseline model
(1.3) without interaction terms. The PV penetration elasticity of volatility of net electricity load
is positive and statistically significant at 5%. The predicted coefficient implies that if the level of
PV penetration in a certain area were to increase by 100%, the volatility of net electricity load on
that transformer would be expected to increase by 3%, ceteris paribus.
The estimated coefficients for average temperature and wind speed are both found to be
statistically significant at 1%. The average temperature has a positive effect on net load volatility
with a 1 °C increase in average temperature increasing the volatility by 5.2%, holding other
variables constant. The fluctuation in electricity demand by households when temperatures are
high, resulting from increased use of air conditioning, for example, creates instability in
electricity use. Conversely, lower temperatures result in a more balanced net electricity load
throughout the day due to a decreased inclination to turn on fans or air conditioning.
The estimated coefficient of wind speed is equal to -0.018 and found to be statistically significant
in all specifications. For every one mile per hour increase in average wind speed, the volatility of
net load is predicated to decrease by 1.8%. Higher and sustained wind speed makes the weather
slightly cooler reducing the need to switch on and off cooling appliances.
Both dummy variables, day/night and summer/winter dummy, are significant at 1%. The positive
coefficient on the day/night dummy variable suggests that net load volatility is higher during the
day than during the evening. This result is clearly shown in figure B.8 which depicts higher
distributions in net load volatility during the daytime. The results in column (1) in table A.3 also
indicate that the volatility of net electricity load is lower during the summer than during the
winter months. Since it is generally sunnier and hotter in the summer, net electricity load is more
consistent and less variable.
16
Column (2) in table A.3 illustrates the impact of customer mix and PV penetration on load
volatility. The coefficient of the interaction term is positive and statistically significant, while the
coefficient of the log of PV penetration is negative and statistically significant. This indicates
that higher PV penetration increases the volatility of net load in residential-concentrated areas
while reducing the volatility in commercial- and industrial-concentrated areas.
Figure B.9 displays annual net load profiles of four different transformers with comparable PV
penetration levels.13 Figure B.9a and figure B.9b depict patterns similar to those observed in
figure B.2 and figure B.3, having slightly lower variation in daytime load due to lower levels of
PV penetration. When comparing figure B.9c and figure B.4, it is observed that figure B.9c
exhibits smaller minimum and higher average net load during the day. This difference is due
primarily to the impact of excess power generation on weekends, as the transformer in figure
B.9c has a smaller number of behind-the-meter rooftop PV installations.
Figure B.9d illustrates the annual net load profile of a transformer consisting of a mixture of
residential and commercial customers (77% residential with several medium-sized commercial
businesses). The average net load of this transformer drops marginally during the daytime,
altering the shape of the load curve. As before, the daytime variability in net load is observed to
have increased over time. However, this increase is less pronounced than was seen in more
residential-concentrated (figure B.9a), commercial-concentrated (figure B.9b) or industrial-
concentrated areas (figure B.9c). Again, it is clearly observable that customer-mix is a key factor
influencing net load behavior in each area. During daytime hours, excess electricity generated by
rooftop solar PV systems is in turn used by the transformer’s commercial customers. The
dynamic between residential and commercial demand patterns greatly reduces back-feeding and
other issues resulting from high variability in net load.
Several existing studies have attempted to determine how heterogeneity in consumer
characteristics and behaviors affect electricity consumption patterns using disaggregated data
Espinoza et al. (2005); Verdú et al. (2006); Lam et al. (2008); Widén and Wäckelgård (2010);
13 These 4 sample transformers have less PV penetration level than figure B.2, B.3, and B.4. Figure B.9a consists of
98% residential customers while figure B.9b comprises only 2% residential customers with two large commercial
businesses. The majority of customers in the transformer in figure B.9c are commercial with two large industrial
businesses and only 4% residential consumers. Figure B.9d consists of a good mix of residential and commercial
with percentage of residential to total equals 77%.
17
Chicco (2012); Flath et al. (2012); Vassileva et al. (2012); Yang et al. (2013); Albert and
Rajagopal (2013). Electric utilities are responsible for supplying power to a wide variety of
different customers. It has been found that electricity consumption patterns are generally similar
within a customer class while differing between classes (Chow et al. 2005). Our results,
therefore, suggest that it is vital for utilities to distinguish load behavior of each customer class
separately, especially when taking into account the impact of a rapid flux in distributed
generations as seen with rooftop PV systems. Understanding the load patterns characterizing
each customer class is of great value to a utility, not only enhancing their ability to better
distribute power generation, but also targeting consumers with appropriate incentive programs.
Column (3) in table A.3 includes the interaction effect of PV penetration and the fluctuation of
solar resource. The main effect of solar irradiance is negative and statistically significant,
implying that when PV penetration is at 0%, as solar irradiance increases the variation in net load
decreases. The coefficient estimate on log of PV penetration becomes insignificant in this
specification, indicating that PV penetration has no effect on the grid at nighttime. The positive
coefficient on the interaction term between log of PV penetration and solar irradiance, however,
suggests that conditional on the level of PV penetration, net load becomes less fluctuating when
it is sunny and there is less cloud in the sky. Given high levels of PV penetration, higher solar
irradiance implies that rooftop PV systems generate higher and more consistent power output
which then results in less variation in net electricity load.
Column (4) in table A.3 provides estimates of the interaction term between temperature and
humidity. The negative coefficient on the interaction term suggests that when humidity is high,
increasing temperatures still result in an increase in volatility of net electricity load, albeit at a
diminishing rate. For example, an increase in temperature from 28 °C to 30 °C will have a
smaller impact on people’s electricity consumption decisions when humidity is already high.
Intuitively, this can be attributed to the fact that most consumers will have already turned on
their air condition when the temperature was at 28 °C.
1.5.1 Additional Result
To capture the effect of a rapid increase in solar penetration, we further evaluate the same
analysis using data only on daytime period (9:00 am to 5:00 pm). Table A.4 reports regression
results from the daytime analysis. From the baseline model, the coefficient on log of PV
18
penetration in column (1) in table A.4 is much greater in magnitude comparing to the same effect
in table A.3. When we consider only daytime, a 100% increase in PV penetration leads to a 9.5%
increase in net load volatility. The positive effect of temperature on net load variation, on the
other hand, decreases from 5.2% to 2.2% with a 1 °C increase in average temperature. The
diversity of customer types on an area also has a greater impact on net load volatility as shown in
column (2).
1.6 Conclusion & Discussion
The unprecedented growth in solar PV adoptions over the past few years has resulted in
remarkably high levels of PV penetration in Oahu, Hawaiʻi. Given this rapid increase in PV
installations, along with the intermittent nature of solar resource itself, the ability to integrate a
vast amount of behind-the-meter rooftop PV systems into the grid has become a costly and
perplexing proposition. This study endeavors to address the underlying determinants of
variability within net electricity load, specifically in light of increasing levels of solar saturation
in Oahu.
Using standard deviation of electricity net load as representative of load volatility, we find that
net load becomes increasingly more volatile as the percentage of PV penetration rises. This
impact, however, is not present at nighttime, implying that consumption behavior of electricity
consumers may remain unaffected by the rapid growth in PV installations. We further assess the
impact of customer diversity on net load volatility and find that customer mix is the key driver
affecting the behavior of electricity net load on each distribution transformer. With an
accelerated increase in solar penetration, the dynamic of electricity consumption behavior
between different types of consumers can essentially lessen problems following higher
fluctuations in electricity net load resulting from variability in solar power output.
From the utility perspective, understanding how net load changes following the rapid increase in
PV installations is crucial. With a more disaggregated data set, electricity consumption patterns
can be used to assess related policies and identify better pricing structures. Finally, the utility can
reduce costs associated with integrating more behind-the-meter solar systems into the electric
grid by fashioning appropriate incentive programs to attract the optimal mix of consumers.
19
CHAPTER 2
Evolution of Residential Solar Adoption in Oahu, Hawaiʻi
2.1 Introduction
Innovation diffusion theory studies the stages underlying the adoption of innovations, the process
of adoption decision-making, and the characteristics of adopters (Rogers 2010). Perception of a
technology, generally based on the characteristics of that technology and people’s level of
awareness, plays an essential role in adoption decisions and, by implication, affects the speed of
diffusion. The awareness stage in the innovation-decision process represents the point at which
adopters gain a full understanding of a technologies’ attributes, and are therefore able to progress
to the decision-making stage. Innovation diffusion is facilitated through various communication
networks over time, influencing the rate of adoption, the innovation decision process, attributes
of new technologies, and the role of change agents. Rogers (2010) divides adopters into five
categories on the basis of their behavior, attitudes, values, personality and the timing of adoption.
These categories are: innovators (2.5% of adopters); early adopters (12.5% of adopters), early
majority (35% of adopters); late majority (35% of adopters); and laggards (15% of adopters).
Under Roger’s classification of adopters, Hawaiʻi households with solar photovoltaic (PV) may
be best categorized as either innovators or early adopters. However, given the rapid evolution of
the Hawaiʻi PV market, which presently has the highest PV penetration rate in the nation, one
might surmise that Hawaiʻi is progressing past the early adopter phase.14 Hawaiʻi, therefore,
serves as a case study for other states presently lagging behind it in the rate of solar adoption. As
the rate of solar adoption increases in other states, Hawaiʻi’s experiences will provide valuable
insight into the unique barriers and challenges inhibiting PV uptake. It is therefore vital that we
gain a better understanding of the adoption process for solar technology, and how it is evolving
over time.
The primary objective of this study is to examine adoption trends and characteristics of PV
adopters on Oahu, Hawaiʻi. We describe both the general attributes characterizing PV adopters
14 As of January 2016, 17% of Oahu customers have had PV installed and 32% of single-family homes on Oahu
have installed solar PV.
20
and the evolution of solar installation trends over the years. By quantifying the various factors
influencing consumers’ solar PV adoption decisions, one is able to improve the efficacy of solar-
related incentives and policies.
Identification of characteristics differentiating PV from non-PV households is done using
detailed information on the time of installation, enabling one to observe the evolution of PV
adopters over time. Homes having PV installations were found to be newer, larger, more energy
efficient, and less costly per square foot than those without a PV installation. The analysis also
revealed that early PV adopters, defined as those installing PV systems before 2012, were
generally older, wealthier, more likely to own their own home, and had higher levels of
educational attainment than did their contemporary counterparts.
To investigate the likelihood of solar adoption among residential single-family households, a
logistic regression model incorporating household consumption level, solar resource availability,
and demographic and housing characteristic information was developed. Empirical results
derived from this model align with descriptive evidence, demonstrating that those living in
larger, newer and less expensive (on a square foot basis) homes were more likely to install solar
PV. Based on demographic information at a census block group level, we find that areas with a
smaller household size, lower median age, higher levels of education, and higher median income
were more likely to have significant solar PV adoption.
The remainder of the paper is organized as follows. Section 2.2 reviews literature related to this
study. Section 2.3 introduces the proprietary dataset underlying the presented analysis and
describes how each variable was processed and summary statistics were developed. Section 2.4
presents descriptive evidence for the data. Section 2.5 details the econometric methodology used
in the analysis, with estimation results reported in Section 2.6. Finally, Section 2.7 offers
concluding remarks and additional discussion of the results.
2.2 Literature Review
There exists a growing body of literature examining the influence of socioeconomic and
customer characteristics and upon the likelihood of solar PV adoption (Keirstead 2007; Rothfield
2010; Kwan 2012; Mills and Schleich 2012; Rai and McAndrews 2012; Balcombe et al. 2013;
Rai and Sigrin, 2013; Langheim et al., 2014; Chernyakhovskiy 2015; Graziano and Gillingham
21
2015). Utilizing zip code level data, Kwan (2012), analyzes the link between a variety of factors
and the distribution of residential solar PV installation. The author found that the level of
available solar resource was the most important factor influencing residential installation, with
higher levels of solar insolation corresponding with increased PV penetration. Other factors
exerting a positive influence on PV adoption include electricity prices, the availability of
financial incentives, and median home values. The paper concluded that the likelihood of PV
adoption was highest among certain groups – namely college educated individuals between 25
and 55 years old with a median income between $25,000 and $100,000 a year.
A number of other studies have employed disaggregated household-level data to identify the
relationship between household characteristics and solar PV adoption. Keirstead (2007), utilizing
demographic information captured in questionnaire response of 91 PV households in the UK,
found that income, education, and homeownership were the primary predictors of solar adoption.
A similar study by Rai and McAndrews (2012) analyzed the socio-demographics of 365 PV
households in Texas using data obtained via a household survey. The survey results found that
residential PV adopters had higher income levels, and were more highly educated than the
average Texas resident.
Although these prior studies assessed the influence of demographics and housing characteristics
on residential solar PV adoption, they failed to analyze the evolution of solar adoption trends
over time. Moreover, their conclusions were based on relatively small samples, which offered no
comparison between PV and non-PV households. The present study aims to fill this literature
gap by analyzing the evolution of PV and non-PV household demographics in order to determine
whether the prototypical solar household has changed over time.
2.3 Data Summary
2.3.1 Data sources
Consumption and PV Installation
Data used in this study were gathered from several different sources. The first dataset consists of
proprietary billing information for 4,047 residential customers covering the period from January
2000 to May 2016. It was made available to the University of Hawaiʻi Research Organization
22
(UHERO) under a confidentiality agreement with Hawaiian Electric Company (HECO), the sole
electric utility on Oahu. The sample was generated by randomly selecting customers using a
random sampling procedure which is explained in detail in Appendix E. Of these customers,
2,490 installed solar PV systems under the Net Energy Metering (NEM) program between
February 2003 and May 2016.
Energy usage metrics and customer information, derived from two ancillary sources, were joined
on the basis of customer installation numbers to form the overall dataset. The first component
dataset contained monthly electricity consumption data in kilowatt-hours (kWh), account and
meter numbers, and meter read dates. Billing data for non-PV customers consisted of actual
monthly electricity consumption. For PV customers, electricity consumption was calculated as
the difference between the amount of electricity delivered from the grid, and the excess
electricity generated by rooftop solar and exported back to the grid by the customer. The second
dataset contained service address, electric distribution zone, an indicator for whether a household
had a solar hot water heater installed, and solar PV installation information (PV system capacity
in kilowatts (kW) and date of installation).
Solar Irradiance
The second dataset, created from two ancillary sources, consists of monthly global horizontal
solar irradiance data (GHI) in Watt/m2 (W/m2).15 The first source, which was provided by AWS
Truepower, LLC and HECO, contains detailed GHI data from over 900 gridded
latitude/longitude points across the island of Oahu. The GHI data points cover the period from
January 2013 through May 2016. The second source was drawn from Clean Power Research’s
SolarAnywhere® PV Power Map and covers the period between 2001 and 2013. It includes
hourly GHI data on a 10-by-10 kilometer (km) tile. For the purposes of this study, 20 tiles
encompassing different parts of Oahu were analyzed.
Housing Characteristics
The third dataset contains housing characteristics, Tax Map Key (TMK) separation, and building
permit information. Housing metadata was obtained from the Real Property Assessment
Division, Department of Budget and Fiscal Services, City and County of Honolulu. For each
15 Global Horizontal Irradiance (GHI) is the solar insolation received by a fixed flat horizontal surface, representing
in the unit of W/m2.
23
household in the sample, the dataset provides information on total property assessed value ($),
types of housing occupancy (single-family and apartment), dwelling size in square footage (sqft),
year built, and number of bedrooms, bathrooms and half-baths. The TMK separation and
building permit details were gathered from the Department of Planning and Permitting (DPP)
and consist of census tract, census block group (CBG), and the total accepted value of solar PV
installations ($) for customers with solar PV.
Census Details
The fourth and final dataset consists of census information obtained from the American
Community Survey (ACS). It includes data elements reported at the CBG level including an
average household size of occupied housing units, a percentage of population 25 years and over
having a college degree or higher, a percentage of owner-occupied homes, median income ($)
and median age.
2.3.2 Data Processing
The aforementioned datasets required considerable manipulation before they could be leveraged
in our analysis. The ensuing section provides a summary of how these datasets were merged and
filtered prior to analysis.
The consumption and PV installation dataset initially contained records for 5,500 residential
customers. Using customer service address information as a key, data elements pertaining to
housing characteristics and building permit information were joined to create a complete
customer information dataset. While generating this dataset it was decided that certain customer
accounts should be excluded should they meet any of the following conditions:
customers for which the name on the property and the HECO account did not match;
customers who had no property information on the Real Property Assessment Division
website; or
customers whose PV statuses under their HECO account and DPP did not match (e.g.,
customers appearing to have solar installation information on DPP but who were not
classified as PV customers within the HECO database).
A total of 695 customers were excluded from the final population based on the above criteria.
24
Next, households residing in apartment buildings were identified using types of housing
occupancy data. It was discovered that 750 households in the initial sample dataset resided in
apartment buildings without solar PV. After excluding these households from the sample, there
were 4,055 residential single-family households remaining in the sample dataset.
Meter read dates were used to determine the month and year for which billing data were to apply.
The logic for this process is as follows: if the meter read date fell between the 1st and 15th of a
given month, then the consumption data reported was assumed to be for the previous month; if
the meter date fell between the 16th and the end of the month, then the consumption data was
assumed to be for that month.
Solar irradiance of PV households in our sample was derived from two distinct data sources. The
estimated monthly solar irradiance for households during the years 2001 to 2012 was sourced
from the publically available SolarAnywhere database, while internal HECO data (AWS) was
utilized for the years 2013 through 2016.
The AWS data covering the latter years in our study consists of measures taken for over 900 1x1
km grid points comprising the island of Oahu. Each household was first assigned the closest
latitude/longitude grid point based on their address using Google Earth as shown in figure D.1.
This mapping produces 216 distinct grid points, for which we query solar irradiance from the
AWS data source. Estimated monthly solar irradiance for the period between 2013 and 2016 was
then determined for each of these grid points, which were subsequently joined to the customer
dataset.
Determination of solar irradiance for the years 2001 through 2012 was performed using hourly
GHI data from the SolarAnywhere dataset. From this data source, we derived total monthly GHI
for each of twenty different tiles which are illustrated in figure D.2. Each customer was then
assigned a tile number (1-20) corresponding to their geographic location.
The SolarAnywhere and AWS datasets overlap in the year 2013. When comparing the values in
each dataset for this overlapping year, we identified inconsistencies between the two measures as
shown in figure D.3. It was, therefore, necessary to adjust the SolarAnywhere observations from
2001-2012 before it was combined with the AWS data in order to achieve a consistent measure
of solar irradiance over the time period considered in the analysis. This was done by first pairing
25
the AWS identifier (1-216) for each household with their corresponding SolarAnywhere tile
number (1-20).
Let AWSit be the estimated monthly solar irradiance of grid point i for month t, while SAit is the
estimated monthly solar irradiance of tile i for month t. Then let AWS-SAit be the estimated
monthly solar irradiance on a combination of the two dataset on a location i at month t. From this
process, we derive distinct 259 grid-tile combinations for customers belonging to our sample
population.
For each of these aforementioned 259 combinations, GHI data is used to calculate an adjustment
factor for each month in the overlapping year (2013) as follows:
AWSit = SAit*it (2.1)
where it is an adjustment factor of a combination i at month t. An overall adjustment factor, i, is
then calculated as the average of these 12 monthly adjustment factors. This overall adjustment
factor is then used to scale SolarAnywhere GHI data for the years 2001-2012 to arrive at a
consistent measure of solar irradiance across the study period as follows:
Adjusted-SIit = SAit*i (2.2)
where Adjusted-SIit is the adjusted values of SAit for tile i during month t. Applying this process,
we obtain monthly solar irradiance (W/m2) for each PV household.
For each PV household, we also determine whether additional PV systems had been added to
their accounts during the observed study period. The initial dataset only indicated the total (i.e.,
current) size of PV systems and the date on which the most recent PV system was installed.
Information detailing the number of additional systems, along with their size and date of
installation, was added using HECO’s internal data portal. From this process, we identified 397
PV accounts with at least one additional PV system installed after the initial PV installation.
Lastly, in order to mitigate the presence of measurement error in monthly reported consumption,
an additional eight accounts were excluded from the study sample. Reasons for exclusion
included unexplained spikes in consumption profiles, prolonged periods of inactivity, and
26
negative or near-zero gross consumption. The complete set of criteria governing exclusion of
specific accounts is as follows:
customer accounts whose gross consumption was negative during any of study months;
customer accounts whose information did not report a solar PV installation, although
their net consumption profiles indicated the presence of a system, having negative
measures for certain months; or
customer accounts exhibiting unusual patterns and/or inconsistent data points. The mean
and standard deviation of monthly consumption were calculated for each customer.
Accounts containing data points exceeding three standard deviations from the mean (>3σ
from the mean) were deleted.
Following this exclusion process, 4,047 residential single-family customer accounts were
ultimately selected for use in the study.
2.3.3 Summary Statistics
The final study sample consisted of 4,047 residential single-family households, 2,490 of which
had installed rooftop PV. The first PV system in our sample was installed in February 2003,
while the latest was installed in May 2016. PV capacity of these systems ranges from 0.28 kW to
35.90 kW.
Table C.1 provides details of variable summary statistics along with t-tests assessing whether
each variable statistically and significantly differed between the two customer groups. Household
consumption was calculated using monthly usage from 2000 to 2005, excluding observations
after PV installation. This measure represents the “baseline” consumption of households in the
sample without the modifying effect of solar installation. It is observed in Table C.1 that PV
households consume approximately 7.5% more energy than non-PV households on average. The
t-test results indicate a statistically significant difference in mean electricity consumption
between PV and non-PV households. The mean monthly solar resource available to households
is found to be equivalent for the two groups.
Housing characteristic information was obtained at the household level. Statistically significant
differences in the mean values were observed between PV and non-PV households, results of
27
which are reported in Table C.1. Home value per square foot was found to be higher for non-PV
households. However, on average, PV homes were found to be larger, newer and use less
electricity on a per square foot basis.
Demographic variables captured at the CBG level exhibited slight differences in their means
between the two customer groups. However, these differences were not found to be statistically
significant, except in the case of median household income at a 4.3% significance level. This
result implies that PV households tend to be located in areas with higher median income.
2.4 Descriptive Evidence
2.4.1 Solar Adoption Trend
This section describes the trend of solar adoption by examining how solar technology has
diffused over time based on the year of PV installation. Figure D.4 illustrates the number of PV
installations and cumulative PV capacity installed of households in the sample. It is observed
that 16% of PV households in our sample have installed additional PV systems after the initial
PV installation.16 The size of these additional PV systems ranges from 0.31 to 15.4 kW, while
the size of original PV systems varies from 0.28 to 35.9 kW.
Several factors have driven the rapid growth of solar PV in Hawaiʻi. First, the availability of
Hawaiʻi solar tax credits and solar incentive programs have played a major role in encouraging
widespread PV adoption. Coffman et al. (2016) address the effect of solar subsidies on
residential PV installations, concluding that investment in solar PV is an exceptional idea for
Hawaiʻi’s homeowners. They argue that various incentives have made solar PV affordable to
many customers, resulting in a significant increase in solar PV adoption.
Secondly, total solar PV installation costs have fallen dramatically over time. We show the trend
in average installation cost of PV in our sample and the trend in the U.S. PV module price in
figure D.5.17 Average installation cost of PV for each year is calculated using the total values of
solar installation obtained from DPP. Comparing the prices of PV systems installed before 2008
with those installed after 2013, we see that the total cost of PV installation has dropped by
16 The number of additional PV systems ranges from 1 to 4 systems per customer. 17 Source: Average value of PV modules, U.S. Energy Information Administration.
28
approximately one-half, increasing the affordability of solar PV and its competitiveness with
other energy sources.
Declining installation costs and solar-friendly policies implemented in Hawaiʻi have led to
remarkable growth in both the number of rooftop PV installations and their average system size
as shown in figure D.6. A PV system installed after 2013 would cost approximately 50% less
than an equivalent system installed prior to 2008. This drop in installation costs has incentivized
consumers to install larger PV systems. The increase in system sizes has raised an anecdotal
issue of whether there exists an “over-sizing” trend in PV installation amongst residential
customers (i.e., the system size chosen by some consumers may be larger than required to satisfy
their energy demand). Given established PV penetration limits on each electrical circuit in Oahu,
this “over-sizing” of PV systems serves to accelerate the speed at which PV penetration
thresholds are met, thereby reducing opportunities for solar adoption by other households.
We next calculate each PV customer’s percentage of consumption offset by their PV systems.
Figure D.7, which shows the percentage of energy offset, illustrates that most households that
adopted PV before 2012 sized their PV systems to displace less than 100% of the total energy
that they consumed. In contrast, the majority of households adopting PV after 2012 installed
systems that offset, on average, 100% of household consumption demand. Figure D.7, therefore,
shows that “over-sizing” of residential PV systems is not widespread. Rather, the increase in
average PV system size observed can be thought of as a natural progression from the “under-
sized” systems installed by early adopters. This result is not entirely surprising given the higher
PV installation costs in the past, larger-than-necessary PV systems were likely not financially
optimal for most residential households.
2.4.2 Characteristics of Adopters & Non-Adopters
Despite the unprecedented growth in residential solar adoption in Hawaiʻi over the past decade,
little attention has been given to examining the types of consumers likely to place solar PV on
their homes. As more households adopt solar PV, the demographics and housing attributes of
adopters also change. In this section, we examine differences in both demographics and housing
characteristics of residential PV adopters and non-PV households on Oahu, Hawaiʻi. Using
detailed information on the time of solar installation, we are not only able to identify which
29
factors are most predictive for solar adoption, but whether these factors are changing as the
technology evolves over time.
Housing Characteristics
Age of Homes
Since the ideal location for solar PV is on a home’s rooftop, it is critical to consider the roof’s
age and condition before installing a PV system. More recently built homes are generally less
likely to require roof replacements to accommodate rooftop solar PV. As a result, one would
expect consumers residing in newer homes to be more likely to install rooftop PV systems
relative to those in aging homes. Figure D.8a illustrates that, in Hawaiʻi, the majority of PV
adopters live in newer homes. However, we also find that the homes of early adopters
(installation prior to 2007) are typically older than those of both recent PV adopters and non-PV
customers.
Home Values
Home value per square foot is calculated using total assessed property value and home size.
From figure D.8b, we find that on average PV homes cost less per square foot than non-PV
homes.
Home Size
Larger homes generally have more rooftop space, resulting in an increased likelihood of PV
adoption. This is illustrated in figure D.8c, which shows that the homes of PV adopters are larger
than non-adopters’ homes on average.
Consumption Level
In terms of consumption level, we calculate an average “baseline” electricity consumption of
each household in the sample. As can be seen in Figure D.8d, most households that adopted PV
before 2010 consume less electricity on average than both recent PV adopters and non-PV
adopters. However, the trend of solar adoption has been transitioning towards high consumption
households in recent years.
30
Home Energy Intensity
Given that increased dwelling size is highly correlated with higher levels of electricity
consumption, we calculate an energy intensity index based on household electricity use per
square foot.18 We observe in figure D.8e that the homes of most PV adopters are more energy
efficient than those without PV. This supports finding from previous studies that adoption of
solar PV is correlated with investment in energy efficiency measures (Haas et al. 1999; Dato
2015). In other words, PV households are more energy-conscious and likely to conserve
electricity.
Demographics
Age
Although data limitations restrict our ability to determine the exact age of individual PV
households in the present study, census level data nonetheless provides some insight. In figure
D.9a we see that most households installing PV before 2009 were typically located in areas
having higher median age than non-PV households and recent PV adopters in the sample. A
decreasing trend in the median age of PV adopters can be observed in the figure, implying that
PV adoption has been transitioning towards younger age groups.
Income
Due to the high upfront cost of PV, households with greater disposable income and better credit
capacity are more likely to purchase solar PV. In figure D.9b we find that the majority of
households that adopted solar PV before 2012 lived in more affluent areas as compared to recent
PV households and non-PV households. The decrease in the price of PV panels in recent years as
shown in figure D.5 has led to a boom in the residential solar market. However, this growth in
solar PV has not been uniformly distributed across the range of household incomes. Our analysis
of the Hawaiian solar market reveals that the diffusion trend is migrating to areas having lower
median income. This observation implies that there may be fewer barriers to solar adoption
18 Average home energy use per square foot is calculated by dividing each household’s average baseline pre-solar
electricity consumption by home size (sqft).
31
among lower income households than there were in the past, resulting in increased rates of
adoption in this consumer market segment.19
Homeownership
The percentage of owner-occupied housing units exhibits a similar trend to median income. In
figure D.9c, we observe that households that installed PV before 2012 were typically located in
areas with higher percentage of owner-occupied homes. Although the correlation between
owner-occupancy and PV adoption rates has lessened in recent years, it may still represent a
limit to the growth of PV adoption due to their requiring rooftops or outdoor/unshaded spaces.
In Oahu, roughly 42% of properties are renter-occupied and many owner-occupied properties are
located in multi-unit buildings or high-rises where it is technically unfeasible to install solar.20
This poses a particular challenge for solar diffusion growth given that renters do not have the
authority to install solar panels, reducing the number of potential solar PV installation sites.
Furthermore, even for properties where solar installations are technically feasible, rental property
owners are not incentivized to invest in PV since they generally do not bear financial
responsibility for electric bills. When tenants are responsible for paying for their own electricity,
landlords have no incentive to install PV systems. In arrangements where rent is fixed and
includes electricity cost, tenants have little to no incentive to conserve energy should the landlord
elect to install solar. As a result, landlords would bear the risk of paying for excess energy usage
on top of the cost of solar installation, discouraging them from adopting solar PV.
Education
Educational attainment is also an essential factor in determining the likelihood of PV adoption.
Figure D.9d demonstrates that the majority of PV adopters before 2012 resided in more educated
areas. Using educational attainment as a proxy for awareness of technology, highly educated
individuals are more likely to adopt solar as they are generally more knowledgeable and 19 In recent years, solar companies have offered a number of different financing options to prospective customers in
order to help offset the initial cost of solar installation. Potential PV adopters may elect to own their own systems by
buying them outright or borrowing against the value of their property through mortgage refinancing via tax
deductible “green energy” loan programs. For households with less financial liquidity and/or lower credit scores,
leasing options and Power Purchase Agreements are also available and require no large upfront expense. The variety
of financing options along with tax credits and other financial incentives will open the solar market to those with
limited access to capital, including lower income households and renters. 20 2011-2015 American Community Survey 5-Year Estimates
32
environmentally aware. Given the complexity of solar technology, it is not wholly unexpected
that early adopters were more highly educated. The declining trend in educational attainment
among PV households observed in figure D.9d may signify that the educational barrier to solar
technology has lessened as more readily understood information about the technology becomes
available through a variety of channels.
Family Size
We find that family size does not differ among PV and non-PV adopters. For each year of
installation, average household size is roughly identical to those without solar PV as seen in
figure D.9e.
2.5 Structural Model
The objective of this study is to explore the likelihood of households installing solar PV through
an evaluation of the determinants of solar PV adoption among residential households in Oahu,
Hawaiʻi. Towards this end, we develop a logistic regression model for PV technology adoption
wherein households make a decision in accordance with their preferences by maximizing the
utility of their energy consumption subject to limitations on their budget constraints. In
particular, we explore which factors drive household i to install a solar PV system. The model
dependent variable is a binary response, taking on the value of 1 if household i installs PV and 0
otherwise. That is,
Yi = 1 if a household i installs a PV system;
0 not install
The study employs several variables that are hypothesized to affect the likelihood of solar
adoption by residential single-family households. These relevant variables include households’
pre-solar electricity consumption, available solar resources, solar hot water heater (SWH)
installation, housing characteristics, and demographic information.
A household’s mean pre-solar consumption is calculated by averaging their monthly electricity
usage from 2000 to 2005, excluding any post-solar observations. This household average pre-
solar usage represents their baseline household energy demand before PV installation. To
measure available solar resources we use the maximum amount of solar resource available to a
33
given household during the 12-month pre-solar period. The presence of a SWH is captured via an
indicator variable that takes the value of 1 if a household has an installed SWH and 0 otherwise.
Household level characteristic variables include property value per square foot, age of the home
and home size. Demographic characteristics, which are gathered at the Census Block Group
level, include education attainment (percent of the population 25 years-old and over that have a
college degree or higher), average household size of occupied housing units, percentage of
owner-occupied homes, median age and median income.
2.6 Empirical Results
Table C.2 reports the marginal effects of the logit model. We find that a household’s probability
of installing PV increases with their electricity consumption, confirming earlier findings in the
literature that higher energy consumption motivates installation of solar PV systems (Balcombe
et al., 2013). In California, Borenstein (2015) found that solar adoption was most prevalent
amongst the highest electricity users. This finding was due in large part to a steeply-tiered
electricity pricing structure under which sample households faced higher marginal prices at
higher-tiered usage levels. Although Hawaiʻi electricity rates are flat, high prices nonetheless
serve to incentivize households to reduce their electricity costs through solar adoption.
In addition to PV system size, PV energy output is largely determined by the availability of solar
resources at a household’s location. With greater solar resource availability, households can
expect higher energy production, leading to more substantial energy bill savings and a higher
return on investment. As a result, one would anticipate that consumers residing in areas having
greater available solar resources are more likely to invest in solar PV (Kwan 2012; Crago and
Chernyakhovskiy, 2014). However, the results of our study reveal that the amount of solar
resources available to Hawaiʻi households do not significantly impact their likelihood of solar
adoption. This deviation from the results of prior studies is largely due to the uniformity of solar
radiation levels across Oahu.
The results of this study are consistent with the findings of previous literature in consumers’
housing characteristics and demographic information (Keirstead 2007; Rothfield 2010; Leenheer
et al. 2011; Willis et al. 2011; Kwan 2012; Mills and Schleich 2012; Balcombe et al. 2013; Rai
and McAndrews 2012; Rai and Sigrin, 2013; Davidson et al. 2014; Langheim et al., 2014;
34
Chernyakhovskiy 2015; Graziano and Gillingham 2015). We find that housing characteristics,
reported in table C.2, statistically and significantly influence the probability of PV installation.
Empirical results indicate that the likelihood of solar adoption increases amongst individuals
residing in newer, larger and less expensive homes (measured on a per square foot basis). Newer
homes typically have better roof conditions which more easily facilitate PV system installation,
while larger homes are correlated with higher electricity consumption. Although the negative
relationship between home value and the likelihood of solar installation seems counterintuitive,
the result is nonetheless consistent with our descriptive finding shown in figure D.8b that most
PV homes have a lower cost per square foot than non-PV homes.
When considering demographic information, we find that the probability of solar adoption
increases in areas with higher median household income, smaller family size, lower median age,
and greater levels of educational attainment.21 Although previous studies have found that
motivation to adopt solar increases with family size (Keirstead 2007; Balcombe et al. 201), we
find a negative relationship between the probability of solar adoption and the number of
individuals in a household. Although larger households tend to consume more electricity, which
would lead to a higher probability of solar adoption, they are also more likely to be financially
constrained by other household expenses.
Homeownership is found to have an insignificant impact on solar installation despite our initially
predicting that homeownership would be highly predictive for solar adoption. This result may be
due in part to the influence of other attributes, such as age and income, which are highly
correlated with homeownership.
The presence of a SWH is found to have a significant effect on the likelihood of PV adoption,
with households having a SWH being inclined to invest in solar PV.22 This result is consistent
with our earlier finding, discussed in the Descriptive Evidence section, that PV homes are more
energy efficient.
21 The effects of these factors may not be straightforward due to interactions of a range of causal factors. 22 Note that in June 2008, Hawaii enacted legislation requiring SWH to be installed on all single-family new home
construction, with a few exceptions (S. 644, 2008). Due to this building energy code, the presence of SWH
installation may not be a significant indicator for the likelihood of solar adoption in the future. In the study sample,
we find that only 0.6% of PV homes with SWH were built after 2008.
35
2.7 Conclusion and Discussion
As national energy policy initiatives continue the push towards clean energy, exemplified in
Hawaiʻi’s embracing of 100% Renewable Portfolio Standards (RPS), there exists an increased
urgency to judiciously divest from traditional fossil fuel based technologies and re-tool using
renewable resources for distributed generations (DG) and other modern technologies.
Foundational planning models need to be enhanced through the integration of refined behavioral
knowledge in conjunction with physical grid constraints, so as to better support sustainable and
efficient diffusion of distributed PV.
To better support the integration of solar PV and other distributed energy resources, it is crucial
to understand the evolution and diffusion of solar PV technology. By evaluating the trends
underlying solar adoption on Oahu, this study revealed that the likelihood of solar adoption was
greatest in newer, larger, more energy efficient and less expensive (per square foot) homes.
Moreover, households living in areas with higher median household income, having smaller
family size, lower median age, and greater levels of educational attainment were found to be
more likely to install solar PV. We also found that having a SWH was the single strongest
predictor of solar PV adoption among residential single-family households.
Future research opportunities abound, including examining the growing trend in solar PV
adoption among non-residential customers. Beyond a per-kWh energy charge, such non-
residential customers are also subject to demand charges which determine the rate schedule to
which they belong.23 Due to this fundamental difference from residential households, their
motivations for PV adoptions can differ greatly from the factors reported on in this study. For
non-residential customers, PV installation can serve to drastically reduce their peak demand,
given their consumption is typically highest during daytime hours which corresponds with peak
solar PV energy production. As a result, the installation of solar PV lessens the probability of
their switching to higher pricing schedules, further reducing their cost of electricity.
Another topic worthy of further study is the potential of battery storage uptake. The rapid decline
in the cost of battery storage technology combined with changes to existing solar incentive
programs, which limit the amount of energy consumers can export to the grid, significantly
23 Demand charges are typically based on the highest level of electricity demand measured in kW.
36
increase the potential influence of battery storage in the near future. Additionally, since the
impact of battery storage discharge behavior on the electrical grid will likely differ significantly
from that of solar technologies, it is vital to assess how the grid may best leverage increased
distributed solar and battery storage penetration to help meet Hawaiʻi’s 100% RPS goal.
37
CHAPTER 3
Impact of Solar Adoption on Residential Electricity Demand
3.1 Introduction
Hawaiʻi has long struggled to identify practical and effective solutions to the unique challenges
facing its energy industry. These challenges arise in large part due to the state’s heavy reliance
on imported fossil fuels for energy generation and the isolated, self-contained, nature of its
electric grid. As a result, Hawaiʻi electricity prices are significantly higher than the U.S. national
average and, as shown in figure G.1, highly correlated with the price of crude oil.24
Given the high electricity prices in Hawaiʻi, the relative economic benefit derived from solar
photovoltaic (PV) technology is greatly enhanced, leading to a high rate of solar PV adoption.25
This study estimates electricity demand on Oahu, Hawaiʻi, examining not only how electricity
usage is impacted by price variations, but how the installation of solar PV and resulting solar PV
Figure G.2 illustrates the relationship between residential monthly electricity consumption in
kilowatt-hours (kWh) and electricity price for Hawaiian PV and non-PV customers in the study
sample from January 2000 to May 2016. It is observed that following the 2008 oil price shock,
which resulted in a spike in Hawaiʻi electricity rates, average energy demand has been steadily
declining. Clear seasonal patterns in average monthly consumption are observable within both
the PV and non-PV customer groups, although the trend begins to exhibit fluctuations towards
the end of the study period as solar PV penetration rises. Variations in consumption among PV
households between summer and winter months become more pronounced beginning in 2013
when the proportion of PV customers in the study sample exceeded 70%.
One of the most important questions relating to post-solar consumption behavior is whether PV
households consume more electricity following adoption. The intuition that solar PV adoption
results in increased electricity consumption stems from the perception that the marginal cost of
24 Over a 15-year period beginning in January 2000, the price of electricity in Hawaiʻi ranged from a low of 15
cents/kWh in 2003 to a high of 40 cents/kWh in 2008. 25 Cumulative PV installations have risen from under 1 megawatt (MW) installed capacity in 2005 to over 280 MW
in mid-2015, with over 95% being customer-sited installations.
38
electricity produced from solar systems is zero, thereby resulting in increased energy demand
amongst PV adopters. However, households that install rooftop PV systems are typically faced
with high upfront installation costs. The energy payback period and the manner in which the
installation is financed will dictate the true price of solar energy production for a given
household.
This study evaluates whether PV adopters exhibit changes in their energy demand, including
responsiveness to price and weather fluctuations, following installation of PV systems. An initial
examination of pre- and post-installation consumption trends within the sample dataset indicated
that PV households increase their electricity usage by approximately 3% in the first year
following PV adoption, with this growth rate gradually decreasing in ensuing years. Conversely,
non-PV customers exhibited consistently decreasing electricity consumption over the observed
time period. However, this cursory analysis considers PV adopters as a homogenous group.
To more clearly understand the impact of solar adoption on electricity consumption, this study
divides PV households on the basis of their PV sizing decisions. Towards this end, we first
define a set of three distinct PV sizing categories: Net Import, those who “under-sized” their PV
systems; Net Zero, those who sized their PV system to offset roughly 100% of their pre-solar
consumption; and Net Export, those who install “larger than necessary” PV systems. Using this
grouping, we find that the majority of households within the sample dataset fall under the Net
Zero group, with only 2% classified as Net Export households.
Following the division of PV households into distinct categories on the basis of their PV sizing
decisions, it is possible to assess how solar installation influences their electricity consumption
behavior. It was observed that Net Import households decrease consumption by approximately
4% in the first year following PV adoption. Conversely, Net Zero households consume more
energy after PV installation, increasing their electricity consumption by approximately 8% in the
first year following PV adoption. Net Export households exhibit the largest post-installation
increase in consumption, which increases by over 30% in the first year following installation and
by over 50% by the end of the fourth year post-installation.
In order to evaluate the dynamics of electricity demand in PV and non-PV households, an
empirical model was developed in this study. We first measure the “baseline” electricity demand
39
of PV and non-PV households utilizing pre-solar observations from January 2000 to December
2009. Analysis of this data reveals that electricity consumers are price-inelastic. In the baseline
period (2000-2009), the price elasticity of demand is similar for PV and non-PV households,
ranging from -0.14 to -0.10. When considering the previously defined PV sizing categories,
results reveal that in the baseline pre-solar period Net Export households exhibit the largest
response to changes in price, while Net Import households are the most inelastic. Non-PV and
Net Zero households are observed to have similar responsiveness to changes in electricity price.
We next estimate electricity demand utilizing both pre- and post-solar installation data spanning
the entire study period from January 2000 to May 2016. Household responsiveness to price and
weather variations is found to differ before and after installation of solar PV systems. Following
PV installation, household consumption becomes more sensitive to price variation, estimated
between -0.25 and -0.17. Clear differences are also observed between the various PV sizing
groups in both their pre-solar responses to price and the impact of installation on their price
response. Electricity consumption in Net Import and Net Zero households becomes more elastic
to price variations following PV installation. Conversely, Net Export households become less
responsive to price after installation of “over-sized” PV systems. This latter observation is not
entirely surprising when considering that Net Export households typically have an excess of
electricity at the end of each billing period. This natural excess in electricity produced versus
electricity demanded provides them sufficient overhead to alter their consumption without
concern for energy price fluctuations.
Results also demonstrate a statistically significant effect of weather on residential electricity
consumption. Temperature is found to have a strong positive correlation with energy
consumption levels. After solar installation, we find that PV households become more sensitive
to weather variations, especially to changes in temperature. This observation mirrors the earlier
descriptive evidence shown in figure G.2, suggesting increased variations in electricity
consumption between summer and winter months among PV households.
The remainder of the paper is organized as follows. Section 3.2 reviews literature related to this
study. Section 3.3 describes the proprietary dataset underlying the presented analysis and details
how each variable was processed. Section 3.4 and 3.5 introduce PV sizing categories and present
summary statistics and descriptive evidence for the data. Section 3.6 details the econometric
40
methodology used in the analysis, with estimation results reported in Section 3.7. Finally,
Section 3.8 offers concluding remarks and additional discussion of the results.
3.2 Literature Review
There is an extensive literature pertaining to demand for electricity that utilizes a wide variety of
econometric estimation methods including time series analysis, partial adjustment model (PAM),
generalized methods of moments (GMM), and ordinary least square estimation (OLS). Table F.1
presents a summary of the existing electricity demand studies in the literature. There is as yet no
clear consensus as to which methodology is most appropriate for electricity demand analysis.
These studies typically incorporate similar control variables, including income, weather,
demographic and dwelling characteristics, while employing different estimation procedures.
These variations can be attributed to the studies’ differing in their length of time covered by the
sample, demand sectors, types of data, and specification of prices. Despite these underlying
differences, the vast majority of studies find price elasticity of electricity demand to be inelastic.
A common challenge when evaluating the relationship between electricity consumption and
price variations is the endogeneity problem.26 Prior studies have generally assumed residential
households to be price takers since their electricity consumption behavior has little to no effect
on changes in electricity prices (Halvorsen and Larsen 1999; Shi et al. 2012). This study utilizes
disaggregated household-level data and a flat electricity price in Hawaiʻi. Therefore, we can
assume each residential household to be a price taker, thereby avoiding the endogeneity problem
in our electricity demand model.
Within this study, we employ a fixed effects model with the log-log functional form to assess
residential electricity demand. The fixed effects model controls for the impact of weather
variation through the inclusion of temperature, wind speed, and rainfall variables. The impact of
weather on electricity consumption has been widely studied in the literature (Kamerschen and
Porter 2004; Filippini 2011). In the residential sector, several studies have found that temperature
is a major determinant of household electricity demand (Silk and Joutz 1997; Hondroyiannis et
al. 2002). Other climatic variables including wind speed and humidity have been used as
26 Besides price endogeneity, another problem arises since PV installation is endogenous. Due to data limitation,
however, we are not able to find variables that can serve as valid instruments, leading to bias in the price elasticity
estimates.
41
correcting terms for the influence of temperature in energy consumption analyses (Engle et al.
1992; Li and Sailor 1995; Cancelo et al. 2008; Yan 1998).
In addition to the aforementioned electricity demand model, this study also explores whether
solar adoption leads to changes in electricity consumption behavior. A number of previous
studies have referred to such a change in electricity consumption behavior as the solar “rebound”
and “double-dividend” effects (McAllister 2012; Blackburn 2014; Deng and Newton 2016). The
notion of rebound effects has been extensively examined and reviewed in the energy efficiency
literature (Khazzoom 1980; Khazzoom 1987; Greening et al. 2000; Sorrell 2007: Sorrell et al.
2009). The rebound effect as outlined in these studies arises when energy consumption increases
as a result of improvements in energy efficiency. The direct rebound effect can be decomposed
into distinct income and substitution effects (Greening et al. 2000; Gillingham et al. 2015). The
income effect reflects the decrease in the cost of energy services leading to an increase in
households’ real income and increased consumption of alternative goods as a result of energy
efficiency improvements. The substitution effect captures the increase in energy consumption in
response to a change in relative prices. Conversely, the “double-dividend” effect leads to
increased conservation following adoption of energy efficiency measures.
Unlike other energy efficient appliances, PV systems are not energy consuming devices.
However, solar PV systems can considerably reduce electricity costs, theoretically incentivizing
households to consume more energy. There is as yet no clear consensus in the literature
regarding how the adoption of PV alters household consumption behavior. Employing
questionnaire data, Keirstead (2007) finds that there exists a solar “double-dividend” effect
among residential households following installation of PV systems. The author also finds that PV
adoption significantly improved awareness of both electricity consumption and generation.
However, this conclusion is drawn based on self-reported information and could, therefore, be
misleading.
Several studies have examined the role that pre-solar usage plays in predicting the effects of
solar adoption. Haas et al. (1999) find that solar adoption triggers increased levels of energy
conservation among high electricity consumers in Austria. High energy users in the
aforementioned study decreased their consumption after PV installation, while low energy users
showed a slight increase in energy demanded. Blackburn (2014) examined post-solar
42
consumption behavior and installation experience via survey and consumption data of residential
households in Texas, finding significant solar “rebound” and “double-dividend” effects arising
after PV installation.
In addition, McAllister (2012) leverages consumption and installation data of 5,243 Californian
households with solar PV to assess the impact of installing a solar system on electricity
consumption. The author examines patterns for system sizing and theorizes that grouping PV
customers on the basis of their pre-solar energy use would lead to a better understanding of post-
solar consumption behavior. The results of McAllister’s study show that the majority of PV
systems in the sample dataset were sized to offset approximately 20% to 80% of households’
total energy demand. Only 10% of the observed households were found to size their PV systems
to displace more than 100% of their pre-solar energy consumption. The author employs the
sizing categorization to evaluate the correlation between sizing decision and post-solar
consumption. The results indicated that PV households with “under-sized” systems relative to
their pre-solar usage tend to demonstrate decreased consumption following installation, whereas
those with larger systems were more likely to increase their level of consumption.
Although these prior studies revealed changes in consumption patterns pre- and post-solar
adoption, no comparison between PV and non-PV adopters was undertaken. Our study aims to
fill this gap in the literature by not only comparing energy usage patterns between pre- and post-
solar installation but also comparing consumption among PV and non-PV households.
3.3 Data Summary
3.3.1 Data Processing
In this study, we employ the same underlying data set from Hawaiian Electric Company (HECO)
utilized in Chapter 2 while incorporating additional electricity price and climatic variables.
Following Ito (2014), it is hypothesized that consumers respond to average price.27 Average
residential electricity price in this study was provided by the U.S. Energy Information
Administration (EIA) and Department of Business, Economic Development and Tourism
(DBEDT). To adjust the electricity price for inflation, the nominal electricity price was divided
27 Given the flat electricity pricing structure in Hawaii, marginal price is equal to the average price for consumers.
43
by the Consumer Price Index (CPI) for all urban consumers (all items) and then multiplied by the
annual average of 2015.28 This adjustment has the effect of normalizing all electricity prices to
2015 dollar values. To account for the effect of weather variation on electricity consumption, the
monthly maximum temperature (Fahrenheit), average wind speed (miles per hour) and total
precipitation (inches) were obtained from the National Weather Service (NWS) for Honolulu
International Airport weather station.
We also determine, for each PV household, whether additional PV systems have been added to
their accounts during the observed time period. It is crucial to validate whether PV customers
have additional installed systems in order to accurately calculate their gross electricity
consumption and determine their appropriate sizing category. Additional information detailing
the number of add-on systems, their size, and the date of installation for each additional system
was added using HECO’s internal data portal. From this process, 397 PV accounts were
identified which had installed at least one additional PV system after the initial PV installation.
Such accounts were excluded from our analysis, leaving 2,093 PV and 1,557 non-PV households
remaining in the sample data set.29
Next, monthly solar electricity produced by rooftop PV is estimated. Let Iit be the solar
irradiance measured at a household i at time t (Watt/m2) and Si be a PV system size of a
household i. The estimated monthly solar electricity produced by a rooftop PV for a household i
at time t (Eit) is calculated as:
Eit = Iit* Si (3.1)
To accurately estimate PV energy output over the course of a solar system’s lifespan, we apply a
0.06% degradation rate per month to the estimated PV energy production (Jordan and Kurtz
2013). Let α be the number of months after the month of PV installation, then the degraded PV
energy output of a household i at time t (DEit) is calculated as:
DEit = Eit*(1.0006)-α (3.2)
28 Source: Consumer Price Index for All Urban Consumers: All items in Honolulu, HI (MSA), Federal Reserve
Bank of St.Louis. 29 PV households with additional PV systems installed were excluded from this study to assure that none of the PV
households have transitioned from one PV sizing group to another during the study period.
44
In order to calculate gross electricity consumption, it is necessary to identify the month in which
the new PV panels become operational. The installation date referenced in the dataset refers to
the date of approval of the interconnection agreement submitted by the customer and does not
always represent the actual date of installation. To mitigate this issue, we identify the first
subsequent month in which a significant reduction in net monthly consumption is detected.
These observations are then used to revise the installation date of each PV customer accordingly.
Gross electricity consumption is calculated by adding the degraded monthly PV energy output
(DEit) to net electricity consumption:
GkWhit = NkWhit + DEit (3.3)
where GkWhit and NkWhit are gross and net electricity consumption of a household i at time t,
respectively. Although we previously adjusted the actual installation date to better reflect the
time of installation, three additional observations – one month before, one month after and the
estimated actual month of installation – are excluded for each customer to ensure clean pre- and
post-solar consumption measurements.
3.3.2 Summary Statistics
Two distinct study datasets, baseline and overall, are considered in the analysis. The baseline
dataset spans the period from January 2000 to December 2009 and excludes observations
occurring after PV installation for each household. The overall dataset consists of all
observations, both pre- and post-solar, from January 2000 to May 2016. Analysis of the baseline
dataset is used to illustrate the starting point of households in each customer group, whereas
analysis of the overall dataset enables us to assess the impact of PV adoption on household
Structural time series model, and TVP= Time varying parameter approach b This table reports short-run residential price elasticities of electricity demand only.
c GCC stands for Gulf Cooperation Council which includes Saudi Arabia, United Arab Emirates, Kuwait, Oman, Bahrain, and Qatar.
d 24 OECD countries include Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Hungary, Ireland, Italy,
Japan, South Korea, Luxembourg, New Zealand, Norway, Portugal, Spain, Sweden, Switzerland, Turkey, UK, and USA.
81
Table F.2: Summary Statistics of Monthly Electricity Usage – PV & No PV